Research map / Label-efficient and robust learning / OOD and anomaly detection
EDGE: Unknown-aware Multi-label Learning by Energy Distribution Gap Expansion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EDGE, an unknown-aware multi-label OOD detection framework that uses BCE plus outlier-exposure losses to widen the JointEnergy gap between tail ID samples and OOD samples, with a feature-based OE dataset selection step.
Problem
Multi-label out-of-distribution detection, where energy-based JointEnergy scores become imbalanced: tail ID samples (only minority-class labels) get energy scores too close to OOD samples, and imbalanced multi-label losses hurt OOD detection.
Approach
Separately optimizes energy for ID and outlier-exposure samples and adds an energy-gap loss that pushes the bottom-k ID JointEnergy scores above OE scores by a margin, plus a singular-value-based dilation distance to choose informative OE datasets.
Key results
- On MS-COCO with ResNet-50, EDGE (JointEnergy inference) reaches FPR95 0.01 on LSUN, 3.09 on Texture and 18.45 on iSUN, versus 0.48, 7.20 and 26.05 for BCE with JointEnergy.
- Ablation on MS-COCO shows the full EDGE loss attains mAP 73.50, compared with 73.79 for BCE alone, while the energy-gap term alone drops mAP to 57.98.
- OE selection on NUS-WIDE picks Img-20-R (lowest dilation distance 13.61) as the best OE set, giving average FPR95 6.45 versus 35.49 for Texture as OE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with EDGE and use JointEnergy inference instead of BCE training with JointEnergy inference | ResNet-50 multi-label OOD detection · MS-COCO → Texture | FPR95: 7.20 -> 3.09 | Helped |
| Add the confidence loss to BCE training (ablation) | ResNet-50 multi-label OOD detection · MS-COCO → Texture | FPR95: 7.20 -> 5.38 | Helped |
| Add the energy-gap loss to BCE training (ablation) | ResNet-50 multi-label OOD detection · MS-COCO → Texture | FPR95: 7.20 -> 8.46 | Hurt |
| Add the confidence loss to BCE training (ablation) | ResNet-50 multi-label OOD detection · MS-COCO → iSUN | FPR95: 26.05 -> 20.39 | Helped |
| Add the energy-gap loss to BCE training (ablation) | ResNet-50 multi-label OOD detection · MS-COCO → iSUN | FPR95: 26.05 -> 43.14 | Hurt |
| Add both confidence and energy-gap losses to BCE training (ablation) | ResNet-50 multi-label OOD detection · MS-COCO → Texture | FPR95: 7.20 -> 3.09 | Helped |
| Add both confidence and energy-gap losses to BCE training (ablation) | ResNet-50 multi-label OOD detection · MS-COCO → iSUN | FPR95: 26.05 -> 18.45 | Helped |
| Select Img-20-R rather than Texture as the OE dataset | EDGE on NUS-WIDE · OOD datasets in the NUS-WIDE evaluation (average) | FPR95: 35.49 -> 6.45 | Helped |
Surprising findings
- Imbalanced multi-label losses (ASL, Focal, DB Loss) degrade OOD detection relative to plain BCE, indicating negative transfer.
- Standard OE and EnergyOE sometimes perform worse than post-hoc inference methods in multi-label settings.
Limitations
- OE dataset quality strongly affects performance; the selection relies on a heuristic dilation distance.
- Hyperparameters (alpha, beta, m, k) are tuned over grids and sensitivity is only partially reported in the main text.
- Some OOD datasets overlap with NUS-WIDE classes, requiring substitution with Img-20-R.
Experimental setup
Datasets: PASCAL-VOC MS-COCO NUS-WIDE ImageNet-22K Img-20-G Img-20-R Textures Places50 iSUN iNaturalist LSUN-C SVHN
Benchmarks: PASCAL-VOC MS-COCO NUS-WIDE LSUN-C Textures iSUN Places50 iNaturalist MNIST SVHN Img-20-R
Metrics: FPR95 AUROC AUPR mAP
Baselines: MaxLogit MSP ODIN Energy Score Mahalanobis ReAct ViM ASH MaxEnergy JointEnergy OE EnergyOE
Builds on: JointEnergy Outlier Exposure EnergyOE ViM
Closest papers by meaning
- Can We Ignore Labels in Out of Distribution Detection? (ICLR 2025 · similarity 0.88)
- ML-GOOD: Towards Multi-Label Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.88)
- Rethinking Out-of-Distribution Detection on Imbalanced Data Distribution (NeurIPS 2024 · similarity 0.88)
- A Noisy Elephant in the Room: Is Your Out-of-Distribution Detector Robust to Label Noise? (CVPR 2024 · similarity 0.87)
- Dual Energy-Based Model with Open-World Uncertainty Estimation for Out-of-distribution Detection (CVPR 2025 · similarity 0.87)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.